The early detection of fire and smoke is a significant aspect in the prevention of disasters in smart cities and the development of extensive monitoring systems. In such scenarios, the early response is critical in ensuring the safety and prevention of further damage. The traditional sensor-based approach in the detection of fire and smoke using heat sensors and smoke detectors is prone to several limitations such as delayed response times, increased false alarm rates, and the inability of the system to adjust to changing environmental conditions. To improve the detection of fire and smoke in the context of a smart-city and the development of extensive monitoring systems, a privacy-preserving vision-based framework called FireSmoke-FL is proposed. FireSmoke-FL is a Federated Learning-based framework that integrates an enhanced YOLOv11 model. The model sensitivity is enhanced by incorporating a high-resolution P2 detection head and an attention-refined C2PSA-iEMA module. These improvements aim to enhance the model’s robustness in detecting fire and smoke in the presence of background interference, such as fog, clouds, and varying illumination conditions. To ensure the privacy of data collected from edge devices, such as IoT devices and drones, FireSmoke-FL is designed to enable devices to learn locally without sharing data. The Dynamic Average Fusion Algorithm (DAFA) is adapted to improve the performance of the model through the adaptive selection of the clients based on the quality of the models developed locally. The FireSmoke-FL framework is validated through extensive experiments on the Fire and Smoke dataset and the Indoor Fire Smoke dataset. The results show that the framework achieves 96.7% mAP at 84.7 FPS and 96.5% mAP at 82.4 FPS on the Fire and Smoke dataset and the Indoor Fire Smoke dataset, respectively. These findings indicate that the proposed model possesses accuracy, efficiency, scalability, and privacy protection.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)